Browse Prior Art Database

Algorithm for Setting Markov Model Probabilities

IP.com Disclosure Number: IPCOM000112604D
Original Publication Date: 1994-Jun-01
Included in the Prior Art Database: 2005-Mar-27
Document File: 2 page(s) / 52K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+5]

Abstract

An algorithm is disclosed to set initial probabilities for Markov model training in a speech recognition system.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Algorithm for Setting Markov Model Probabilities

An algorithm is disclosed to set initial probabilities for Markov
model training in a speech recognition system.

      One of the major difficulties encountered in training those
models (i.e., adapting the probabilities to a training set) is to
come up with sensible initial values for the label distribution
probabilities.  The most common method is to initialize the
distributions using random numbers, and do several training runs
starting from different random starting points [1,2].

      The closer the initial values are to the final trained model's
values, the faster they converge to a "well" trained model.  The idea
at the base of the algorithm disclosed here is to use the training
data (sequences of observed acoustic events or "labels") to set up
the initial statistics of the Markov model baseform for each word.

      Each sequence of observed labels is aligned with the baseform
nodes by a simple linear time warping:
    If the baseform has nbNode nodes (from 0 to nbNode-1),
    and the observed label sequence has nbLabel labels (from 0 to
    nbLabel-1),
    then node i is assigned label number j sub c (i)= <<i *
    (nbLabel-1)> over <nbNode-1>> from the
    observed sequence as the "center label".

      The observed label value found for the center label increases
the count for the emission of that label at that node.  In addition,
neighboring nodes also see that label's count in...